Toward coherent space–time mapping of seagrass cover from satellite data: an example of a Mediterranean lagoon

Seagrass meadows are a highly productive and economically important shallow coastal habitat. Their sensitivity to natural and anthropogenic disturbances, combined with their importance for local biodiversity, carbon stocks, and sediment dynamics, motivate a frequent monitoring of their distribution. However, generating time series of seagrass cover from field observations is costly, and mapping methods based on remote sensing require restrictive conditions on seabed visibility, limiting the frequency of observations. In this contribution, we examine the effect of accounting for environmental factors, such as the bathymetry and median grain size (D50) of the substrate as well as the coordinates of known seagrass patches, on the performance of a random forest (RF) classifier used to determine seagrass cover. Using 148 Landsat images of the Venice Lagoon (Italy) between 1999 and 2020, we trained an RF classifier with only spectral features from Landsat images and seagrass surveys from 2002 and 2017. Then, by adding the features above and applying a time-based correction to predictions, we created multiple RF models with different feature combinations. We tested the quality of the resulting seagrass cover predictions from each model against field surveys, showing that bathymetry,D50, and coordinates of known patches exert an influence that is dependent on the training Landsat image and seagrass survey chosen. In models trained on a survey from 2017, where using only spectral features causes predictions to overestimate seagrass surface area, no significant change in model performance was observed. Conversely, in models trained on a survey from 2002, the addition of the out-of-image features and particularly coordinates of known vegetated patches greatly improves the predictive capacity of the model, while still allowing the detection of seagrass beds absent in the reference field survey. Applying a timebased correction eliminates small temporal variations in predictions, improving predictions that performed well before correction. We conclude that accounting for the coordinates of known seagrass patches, together with applying a timebased correction, has the most potential to produce reliable frequent predictions of seagrass cover. While this case study alone is insufficient to explain how geographic location information influences the classification process, we suggest that it is linked to the inherent spatial auto-correlation of seagrass meadow distribution. In the interest of improving remote-sensing classification and particularly to develop our capacity to map vegetation across time, we identify this phenomenon as warranting further research.